Pith. sign in

REVIEW 3 major objections 5 minor 41 references

TBSG-Net: Temporal Bipartite Scene Graph Network for Fine-Grained Video Moment Retrieval

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TBSG-Net claims that encoding how long object interactions last—not just which objects and relations appear—is what enables fine-grained video moment retrieval, setting new state-of-the-art results on three Charades-STA variants.

desk verdict Solid new graph construction for VMR, but the temporal-span evidence doesn't yet rule out a duration shortcut. read the letter →

arxiv 2608.02056 v1 pith:OSAPVUWY submitted 2026-08-03 cs.CV cs.AI

Ji Huang , Yongsheng Dai , Tianyu Ren , Barry Devereux , Hui Wang This is my paper
classification cs.CVcs.AI
keywords videomomentretrievaldynamicscenegraphtemporalbipartitespanencodingproposal-freerelation-awareattentionconvolutionalnetworkzero-shottransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Video moment retrieval asks a model to find the segment of an untrimmed video that matches a natural-language query. Prior proposal-free methods that use static scene graphs capture objects and relations per frame but ignore how those relations develop over time. TBSG-Net instead builds on dynamic scene graphs—tracks of objects and their relations over time—and converts them into temporal bipartite graphs whose relation nodes carry explicit start–end spans. Those spans are used to weight attention and graph convolution. The paper reports consistent gains over prior methods, with the largest improvements on benchmark splits designed to test robustness to shifts in moment duration and location.

What carries the argument

The Temporal Bipartite Scene Graph (TBSG): a bipartite graph with two disjoint node sets—objects and relationships—where edges only connect an object to a relation in which it participates. Each relation node stores a temporal span T_ij = sum over contiguous intervals of (end − start) during which that interaction is detected. The same span enters the architecture twice: as an exponential duration-aware weight in the transformer attention (T_ij = exp(λ·T_ij/T_max)) and as a scaling factor on the adjacency matrix before symmetric-normalized graph convolution. The bipartite structure lets the encoder reason separately over appearance (objects) and semantics (relations) before fusing, while the

What would settle it

Train TBSG-Net on Charades-STA-Len with each relation span replaced by a random duration drawn from the same distribution; if R@1 at IoU=0.7 stays at 25.99, the span signal is not what drives the gain. Also compute the correlation between summed relation spans and target moment durations on the test split; a strong correlation would indicate the anti-bias gains may be inflated by leakage.

Watch

Extended reading notes

Core claim

The central claim is that the missing ingredient for fine-grained video moment retrieval is not more frame-level semantics but an explicit representation of when each object–relation interaction begins, ends, and how long it persists. TBSG-Net constructs a temporal bipartite scene graph from an off-the-shelf dynamic scene graph detector: object nodes and relation nodes form a bipartite graph, and each relation node carries a temporal span computed by merging contiguous detections of the same relation. A duration-aware exponential weight multiplies the attention scores and the graph adjacency, so sustained interactions dominate the encoding. Combining a transformer for global event dependenci

Load-bearing premise

The reported gains rest on the assumption that the offline dynamic scene graph stream provides accurate and temporally consistent object-relation detections, and that the temporal spans computed from those detections are a genuine semantic signal rather than a proxy for the target moment's duration or position.

Editorial extensions

If this is right

  • If the central claim holds, proposal-free video moment retrieval can be improved by adding a lightweight graph stream without changing the decoder.
  • Explicit temporal spans should transfer across datasets whenever a dynamic scene graph detector can be run offline, as the zero-shot ActivityNet experiment indicates.
  • The largest gains at strict IoU=0.7 on the anti-bias splits suggest the span signal is what carries robustness to duration and position shifts.
  • Duration-aware weighting is a general idea: any model with attention over events could incorporate interaction persistence rather than just occurrence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable corollary the paper does not pursue: if T_ij is replaced by the total time an object is visible regardless of relation, the gains should shrink, isolating the relation-specific part of the span signal.
  • On datasets with more than one subject, the paper's single scalar span per relation loses role information; the subject/object role matrices suggested in the future-work section would likely matter more there than on Charades-STA.
  • The zero-shot gap shrinking relative to in-domain results hints that the detector's vocabulary coverage, not the architecture, is the binding constraint; a pseudo-labelling pipeline that expands coverage could recover most of the gap.
  • Because the span is computed per relation pair across the whole clip, the method implicitly assumes reliable object identity tracking; identity switches would corrupt spans, which is worth stress-testing with an identity-tracking oracle.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes TBSG-Net, a proposal-free video moment retrieval method built on dynamic scene graphs (DSGs). Offline STTran generates frame-level DSGs; a TBSG Constructor aggregates them into a bipartite graph with object and relation nodes annotated by scalar temporal spans T_ij; a hybrid Transformer-GCN encoder produces clip-level embeddings that are fused with coarse video features and fed to a UniMD-style decoder. Experiments on Charades-STA, two anti-bias variants (Len, Mom), and zero-shot transfer to ActivityNet Captions report state-of-the-art results, supported by ablations of the DSG stream, the DSG-E module, architecture ordering, time-span encoding, hyperparameters, and upstream noise sensitivity.

Significance. If the claims hold, the paper is a meaningful advance: it is the first DSG-based proposal-free VMR model, it shows consistent gains across four visual backbones and two deliberately biased splits, and it includes an unusually thorough robustness analysis (upstream threshold perturbations, relation-node corruption, loss-weight sensitivity). The code release, parameter studies, and machine-checkable ablations are strengths. However, the central mechanistic claim—that explicit temporal-span encoding of relational structure drives the gains—is not fully established. The randomized-span control in Table X cannot rule out a duration shortcut, and the quality of the DSG inputs that define T_ij is never quantified. The contribution is therefore promising but needs additional evidence before the interpretation can be accepted.

major comments (3)
  1. [IV-E4/Table X] The randomized-time-span ablation does not distinguish semantic temporal structure from a duration shortcut. T_ij is a scalar sum of active interval lengths (Section III-E1), and Eq. (10) injects only this scalar into attention. On Charades-STA-Len and Charades-STA-Mom, the target duration and position distributions are deliberately shifted, so any feature correlated with target duration or position can inflate high-IoU R@1 even without relational semantics. Replacing true durations with independently randomized values changes both the marginal duration distribution and the relation-to-duration assignment, so the large drop in Variant (3) is equally consistent with the model relying on a low-dimensional duration proxy. I request a within-clip permutation control (e.g., permuting T_ij across relation edges inside each clip while preserving graph structure and the marginal distribution) or
  2. [III-D/IV-A] The DSG stream that defines T_ij is produced by a pre-trained STTran model, but the paper reports no accuracy, recall, or noise level for these detections on Charades-STA or on the pseudo-labelled ActivityNet Captions data. Section IV-A states that Charades-STA has scene-graph annotations, so a direct comparison of STTran's DSGs with ground-truth graphs is feasible and would calibrate how much of the improvement comes from relational semantics rather than detection artifacts. This missing measurement also weakens the inference-time robustness claims in Section IV-F4, which assume specific corruption levels without knowing the actual upstream error rate.
  3. [III-E1/E2] The paper repeatedly claims explicit encoding of temporal 'spans' (start and end times), but the method only encodes scalar duration. Eq. (6) defines intervals [t_s, t_e], yet these are immediately aggregated into T_ij = sum_k (t_k^e - t_k^s), and Eq. (10) adds this scalar to the attention logits. Positional information about when a relation begins and ends is discarded. This undercuts the stated distinction from SSGs regarding 'when an interaction begins and how long it lasts' (Section I) and is particularly relevant to the Charades-STA-Mom position-shift benchmark. The authors should either enrich the representation with interval endpoints or temper the 'temporal span' terminology.
minor comments (5)
  1. [IV-C/Table II] The text states that the VGG+GloVe variant 'surpasses all GloVe-based baselines across all metrics', but Table II shows R@5 at IoU=0.5 of 88.27 for TBSG-Net versus 88.39 for CBLN. Please correct the claim or the table.
  2. [IV-B and all result tables] All results are reported as single runs without standard deviations or significance tests. Given the large claims, please provide mean±std over at least three seeds, or state that only single runs were performed.
  3. [III-E] The notation T is overloaded: Eq. (6) defines a set of intervals, while Eq. (10) uses a matrix with entries T_ij, and the duration weighting is written T_ij = exp(...). Please use distinct symbols (e.g., S for the interval set) and define Tmax precisely as the dataset-wide maximum video duration versus the clip duration.
  4. [IV-F/Fig. 5-7] Figures 5-7 show trends without error bars or point-level variability. Adding error bars or reporting the underlying values would strengthen the parameter-study conclusions.
  5. [IV-A/III-E1] The statement that 'the entity Person is the only subject class' in Charades-STA should be verified and, if true, its implications for the generality of the DSG construction should be discussed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: T_ij is an input-derived feature, not a renamed target; no self-citation chain carries the argument.

full rationale

The paper's central derivation is self-contained in the relevant sense. The temporal spans T_ij are computed from the input video's detected Dynamic Scene Graphs (Section III-E1, Eq. 6 and the aggregation T_ij = Σ_k(t_k^e - t_k^s)), not from the retrieval target x=(t_s, t_e). The target boundaries enter only through the supervised classification/regression losses (Eq. 14-16). The duration-aware weighting T_ij = exp(λ·T_ij/T_max) in Eq. (10) uses these input spans as features; λ is a learnable parameter trained end-to-end, not a fitted target renamed as a prediction. The ablation study in Table X tests no, binary, randomized, and full span encodings, so the contribution of temporal spans is empirically evaluated rather than assumed. The 'first DSG-based proposal-free VMR model' claim is a novelty assertion, not a uniqueness theorem or a self-citation. The cited foundations (STTran [17], UniMD [9], PaTF [7], MomentDiff [4]) are independent external works; no load-bearing self-citation appears. The skeptic's duration-shortcut concern on the anti-bias splits is a validity/robustness question, not circularity: a possible empirical correlation between relation durations and target moment duration does not make T_ij definitionally equivalent to the target, and the randomized-span ablation is an empirical control rather than a construction that forces the result.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method introduces no physical entities. The central claim rests on the reliability of the upstream DSG detector, the compatibility of its vocabulary with each dataset, and the modeling choice of representing relation persistence as summed contiguous spans. The main fitted quantity is the learned duration weight lambda, and several hyperparameters are chosen from parameter studies on the evaluation benchmark.

free parameters (5)
  • duration weighting lambda = learned, initialized at 0.1, final value not reported
    Scalar in T_ij = exp(lambda * T_ij / T_max) used in Eq. 10; learned end-to-end on the retrieval loss and controls sensitivity to interaction duration.
  • frames per DSG F = 4
    Selected from parameter study in Fig. 5, which shows best performance at F=4 on Charades-STA; no separate validation split is reported.
  • max objects per DSG o = 25
    Selected from parameter study in Fig. 6a, where performance plateaus near o=25.
  • max relations per DSG r = 30
    Selected from parameter study in Fig. 6b, where performance plateaus near r=30.
  • Transformer and GCN layer counts = 2 Transformer layers, 3 GCN layers
    Selected from grid in Fig. 7 as the configuration with highest R@1 on Charades-STA.
assumptions (3)
  • domain assumption STTran-generated dynamic scene graphs are accurate enough to provide an event-level relational signal on Charades-STA and ActivityNet Captions.
    Section III-D uses STTran as the relational detector and Section IV.D applies pre-trained STTran to ActivityNet. No DSG accuracy or noise level is reported, yet the whole fine-grained stream depends on these detections.
  • domain assumption The object and relation vocabulary of Charades-STA is compatible with the relational detector used to build DSGs.
    Section IV.A says Charades-STA is the only widely used VMR dataset with scene graph annotations, but STTran is trained on Action Genome; the vocabulary overlap is not quantified for Charades-STA, only acknowledged as reduced for ActivityNet in Section IV.D.
  • domain assumption Merging consecutive timestamps into contiguous temporal spans preserves the event structure needed for moment localization.
    Section III-E1 defines T_ij as a sum of contiguous intervals, and this aggregation is a modeling choice. The randomized-span ablation supports duration sensitivity, but not that the merging rule is semantically optimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TBSG-Net: Temporal Bipartite Scene Graph Network for Fine-Grained Video Moment Retrieval." pith.science (2026). https://pith.science/paper/OSAPVUWY

@misc{pith2026260802056,
  author       = {Pith},
  title        = {Pith review of: TBSG-Net: Temporal Bipartite Scene Graph Network for Fine-Grained Video Moment Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSAPVUWY}},
  note         = {Machine review of arXiv:2608.02056}
}
read the original abstract

Recent advances in proposal-free Video Moment Retrieval (VMR) have highlighted the effectiveness of Static Scene Graphs (SSGs). By modeling objects and their relations at the frame level, SSGs enrich retrieval-oriented video representations. However, integrating SSGs into VMR remains constrained by two inherent limitations: (1) Lack of Temporal Dynamics. SSGs fail to model how objects and their relationships evolve over time, leading to the loss of essential temporal dependencies in video representation; and (2) Lack of Explicit Temporal Span Encoding. SSGs do not explicitly encode the duration of relationships, making precise localization challenging. To address these limitations, we propose Temporal Bipartite Scene Graph Network (TBSG-Net)---to the best of our knowledge, the first Dynamic Scene Graph (DSG) based proposal-free VMR model. Specifically, TBSG-Net leverages DSGs to extract event-centric graph representations of the input video, enabling the modeling of object interactions over time and thus addressing limitation (1). These DSGs are then processed by a novel Dynamic Scene Graph Embedding (DSG-E) module to capture both Temporal Span and spatio-temporal information. First, DSG-E utilizes a TBSG Constructor to transform DSGs into TBSGs, explicitly encoding objects, relationships, and time spans to tackle limitation (2). Second, the resultant TBSGs are passed into a hybrid TBSG Encoder that integrates a Transformer variant for global event modeling and a Graph Convolutional Network for detailed relational reasoning, ultimately producing a more comprehensive spatio-temporal representation. Our experiments demonstrate substantial improvements of TBSG-Net over all baselines.

Figures

Figures reproduced from arXiv: 2608.02056 by the authors.

Figure 1
Figure 1. DSG Generation Process: (1) The Spatial Module analyzes objects [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The TBSG-Net architecture follows an encoder-decoder paradigm. A Text Encoder first captures the semantic context of the input queries, while [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. First, we compute the union sets of unique objects [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Demonstration of TBSGs. Nodes are divided into objects and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]
Figure 4
Figure 4. Figure 4: Three configurations of relative positioning in the Transformer-GCN. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Impact of the number of frames per dynamic scene graph on R@1 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Effect of the maximum number of objects o and relationships r per DSG on retrieval performance (R@1 at IoU 0.5 and 0.7) on Charades-STA. (a) R@1, IoU=0.5 (b) R@1, IoU=0.7 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Effect of Transformer and GCN layer counts on Charades-STA. The [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Two Top-1 qualitative comparisons from Charades-STA. The three [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 2 linked inside Pith

  1. [1]

    Tall: Temporal activity lo- calization via language query,

    J. Gao, C. Sun, Z. Yang, and R. Nevatia, “Tall: Temporal activity lo- calization via language query,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 5267–5275

    2017

  2. [2]

    Retrieval-specific view learn- ing for sketch-to-shape retrieval,

    S. Yuan, C. Wen, Y .-S. Liu, and Y . Fang, “Retrieval-specific view learn- ing for sketch-to-shape retrieval,” IEEE Transactions on Multimedia, 2023

    2023

  3. [3]

    Beatdance: A beat-based model-agnostic contrastive learning frame- work for music-dance retrieval,

    K. Yang, X. Zhou, X. Tang, R. Diao, H. Liu, J. He, and Z. Fan, “Beatdance: A beat-based model-agnostic contrastive learning frame- work for music-dance retrieval,” inProceedings of the 2024 International Conference on Multimedia Retrieval, 2024, pp. 11–19

    2024

  4. [4]

    Momentdiff: Generative video moment retrieval from random to real,

    P. Li, C.-W. Xie, H. Xie, L. Zhao, L. Zhang, Y . Zheng, D. Zhao, and Y . Zhang, “Momentdiff: Generative video moment retrieval from random to real,” Advances in neural information processing systems, vol. 36, 2024

    2024

  5. [5]

    Detecting moments and highlights in videos via natural language queries,

    J. Lei, T. L. Berg, and M. Bansal, “Detecting moments and highlights in videos via natural language queries,” Advances in Neural Information Processing Systems, vol. 34, pp. 11 846–11 858, 2021

    2021

  6. [6]

    Learning 2d temporal adjacent networks for moment localization with natural language,

    S. Zhang, H. Peng, J. Fu, and J. Luo, “Learning 2d temporal adjacent networks for moment localization with natural language,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 07, 2020, pp. 12 870–12 877

    2020

  7. [7]

    A parallel transformer framework for video moment retrieval,

    T.-N. Nguyen, Z. Li, Y . Satoshi, J. Liu, and C. Gurrin, “A parallel transformer framework for video moment retrieval,” in Proceedings of the 2024 International Conference on Multimedia Retrieval, 2024, pp. 460–468

    2024

  8. [8]

    Saliency- guided detr for moment retrieval and highlight detection,

    A. Gordeev, V . Dokholyan, I. Tolstykh, and M. Kuprashevich, “Saliency- guided detr for moment retrieval and highlight detection,” arXiv preprint arXiv:2410.01615, 2024

    arXiv 2024

Show all 41 references
  1. [9]

    Unimd: Towards unifying mo- ment retrieval and temporal action detection,

    Y . Zeng, Y . Zhong, C. Feng, and L. Ma, “Unimd: Towards unifying mo- ment retrieval and temporal action detection,” in European Conference on Computer Vision. Springer, 2025, pp. 286–304

    2025

  2. [10]

    Estimating the semantics via sector embedding for image-text retrieval,

    Z. Wang, Z. Gao, M. Han, Y . Yang, and H. T. Shen, “Estimating the semantics via sector embedding for image-text retrieval,” IEEE Transactions on Multimedia, vol. 26, pp. 10 342–10 353, 2024

    2024

  3. [11]

    Semantics disentangling for cross-modal retrieval,

    Z. Wang, X. Xu, J. Wei, N. Xie, Y . Yang, and H. T. Shen, “Semantics disentangling for cross-modal retrieval,” IEEE Transactions on image processing, vol. 33, pp. 2226–2237, 2024

    2024

  4. [12]

    Image retrieval using scene graphs,

    J. Johnson, R. Krishna, M. Stark, L.-J. Li, D. Shamma, M. Bernstein, and L. Fei-Fei, “Image retrieval using scene graphs,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3668–3678

    2015

  5. [13]

    Look twice as much as you say: Scene graph contrastive learning for self- supervised image caption generation,

    C. Zhang, C. Huang, Y . Li, X. Zhang, Y . Ye, and C. Zhang, “Look twice as much as you say: Scene graph contrastive learning for self- supervised image caption generation,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp....

    2022

  6. [14]

    Image-to-image retrieval by learning similarity between scene graphs,

    S. Yoon, W. Y . Kang, S. Jeon, S. Lee, C. Han, J. Park, and E.-S. Kim, “Image-to-image retrieval by learning similarity between scene graphs,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 10 718–10 726

    2021

  7. [15]

    Sgtr: End-to-end scene graph generation with transformer,

    R. Li, S. Zhang, and X. He, “Sgtr: End-to-end scene graph generation with transformer,” in proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 19 486–19 496

    2022

  8. [16]

    Complex relation embedding for scene graph generation,

    Z. Wang, X. Xu, Y . Zhang, Y . Yang, and H. T. Shen, “Complex relation embedding for scene graph generation,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 6, pp. 8321–8335, 2022

    2022

  9. [17]

    Spatial-temporal transformer for dynamic scene graph generation,

    Y . Cong, W. Liao, H. Ackermann, B. Rosenhahn, and M. Y . Yang, “Spatial-temporal transformer for dynamic scene graph generation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 16 372–16 382

    2021

  10. [18]

    Scenegate: Scene-graph based co-attention networks for text visual question an- swering,

    F. Cao, S. Luo, F. Nunez, Z. Wen, J. Poon, and S. C. Han, “Scenegate: Scene-graph based co-attention networks for text visual question an- swering,” Robotics, vol. 12, no. 4, p. 114, 2023

    2023

  11. [19]

    Relational reasoning over spatial- temporal graphs for video summarization,

    W. Zhu, Y . Han, J. Lu, and J. Zhou, “Relational reasoning over spatial- temporal graphs for video summarization,” IEEE Transactions on Image Processing, vol. 31, pp. 3017–3031, 2022

    2022

  12. [20]

    Multi-modal relational graph for cross-modal video moment retrieval,

    Y . Zeng, D. Cao, X. Wei, M. Liu, Z. Zhao, and Z. Qin, “Multi-modal relational graph for cross-modal video moment retrieval,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 2215–2224

    2021

  13. [21]

    Modality-aware heterogeneous graph for joint video moment retrieval and highlight detection,

    R. Wang, J. Feng, F. Zhang, X. Luo, and Y . Luo, “Modality-aware heterogeneous graph for joint video moment retrieval and highlight detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 9, pp. 8896–8911, 2024

    2024

  14. [22]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  15. [23]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  16. [24]

    Quo vadis, action recognition? a new model and the kinetics dataset,

    J. Carreira and A. Zisserman, “Quo vadis, action recognition? a new model and the kinetics dataset,” in proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 6299–6308

    2017

  17. [25]

    Slowfast networks for video recognition,

    C. Feichtenhofer, H. Fan, J. Malik, and K. He, “Slowfast networks for video recognition,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6202–6211

    2019

  18. [26]

    Debug: A dense bottom- up grounding approach for natural language video localization,

    C. Lu, L. Chen, C. Tan, X. Li, and J. Xiao, “Debug: A dense bottom- up grounding approach for natural language video localization,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural L...

    2019

  19. [27]

    Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment,

    D. Zhang, X. Dai, X. Wang, Y .-F. Wang, and L. S. Davis, “Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1247–1257

    2019

  20. [28]

    Dori: Discovering object relationships for moment localiza- tion of a natural language query in a video,

    C. Rodriguez-Opazo, E. Marrese-Taylor, B. Fernando, H. Li, and S. Gould, “Dori: Discovering object relationships for moment localiza- tion of a natural language query in a video,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 1079–1088

    2021

  21. [29]

    Context-aware biaffine localizing network for temporal sentence grounding,

    D. Liu, X. Qu, J. Dong, P. Zhou, Y . Cheng, W. Wei, Z. Xu, and Y . Xie, “Context-aware biaffine localizing network for temporal sentence grounding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 11 235–11 244

    2021

  22. [30]

    Diving into the relations: Lever- aging semantic and visual structures for video moment retrieval,

    Z. Wu, J. Gao, S. Huang, and C. Xu, “Diving into the relations: Lever- aging semantic and visual structures for video moment retrieval,” in 2021 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2021, pp. 1–6

    2021

  23. [31]

    Negative sample matters: A renaissance of metric learning for temporal grounding,

    Z. Wang, L. Wang, T. Wu, T. Li, and G. Wu, “Negative sample matters: A renaissance of metric learning for temporal grounding,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 3, 2022, pp. 2613–2623

    2022

  24. [32]

    Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection,

    Y . Liu, S. Li, Y . Wu, C.-W. Chen, Y . Shan, and X. Qie, “Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 3042–3051

    2022

  25. [33]

    Cross-modal dynamic networks for video moment retrieval with text query,

    G. Wang, X. Xu, F. Shen, H. Lu, Y . Ji, and H. T. Shen, “Cross-modal dynamic networks for video moment retrieval with text query,” IEEE Transactions on Multimedia, vol. 24, pp. 1221–1232, 2022

    2022

  26. [34]

    Query-dependent video representation for moment retrieval and highlight detection,

    W. Moon, S. Hyun, S. Park, D. Park, and J.-P. Heo, “Query-dependent video representation for moment retrieval and highlight detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 23 023–23 033

    2023

  27. [35]

    Unloc: A unified framework for video localization tasks,

    S. Yan, X. Xiong, A. Nagrani, A. Arnab, Z. Wang, W. Ge, D. Ross, and C. Schmid, “Unloc: A unified framework for video localization tasks,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 13 623–13 633

    2023

  28. [36]

    Towards generalisable video moment retrieval: Visual-dynamic injection to image-text pre- training,

    D. Luo, J. Huang, S. Gong, H. Jin, and Y . Liu, “Towards generalisable video moment retrieval: Visual-dynamic injection to image-text pre- training,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23 045–23 055

    2023

  29. [37]

    Background-aware moment detection for video moment retrieval,

    M. Jung, Y . Jang, S. Choi, J. Kim, J.-H. Kim, and B.-T. Zhang, “Background-aware moment detection for video moment retrieval,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 2025, pp. 8586–8596

    2025

  30. [38]

    Object- centric framework for video moment retrieval,

    Z. Li, Y . Wong, S. Yamazaki, J. Liu, and M. Kankanhalli, “Object- centric framework for video moment retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 8, 2026, pp. 6771–6779

    2026

  31. [39]

    Natural language video localization with learnable moment proposals,

    S. Xiao, L. Chen, J. Shao, Y . Zhuang, and J. Xiao, “Natural language video localization with learnable moment proposals,” arXiv preprint arXiv:2109.10678, 2021

    2021 arXiv

  32. [40]

    Fast video moment retrieval,

    J. Gao and C. Xu, “Fast video moment retrieval,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1523–1532

    2021

  33. [41]

    Dense- captioning events in videos,

    R. Krishna, K. Hata, F. Ren, L. Fei-Fei, and J. Carlos Niebles, “Dense- captioning events in videos,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 706–715

    2017

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.